A new inference-time method called SatisDive tries to stop AI image generators from gaming diversity scores by hiding low-quality images in a batch.
Researchers behind a paper posted to arXiv this week argue that current text-to-image systems handle reward (how well an image matches what a user wants) and diversity (how different the images in a batch look from each other) in ways that let one measure mask failures in the other. Some methods score reward and diversity separately. Others fold both into one combined score, which lets a batch look good overall even if individual images are weak, as long as they look different enough from each other. SatisDive instead requires every image to clear a minimum reward floor first, then optimizes for diversity only among the images that pass. In tests on the Pick-a-Pic dataset using FLUX.1-dev and SANA-1.6B as base models, the researchers report SatisDive improved the worst-scoring image's reward by up to 0.43 and 0.70 points respectively over an existing technique called FK steering, at matched diversity scores.
This is a narrow fix, but it names a real failure mode in generative tools: optimizing for an averaged or combined score can quietly hide bad outputs. Anyone who has generated four variations of a prompt and gotten three good images and one dud has run into exactly the problem this paper targets.
It is not a new model or a splashy demo, just a reordering of priorities at inference time, the kind of unglamorous fix that tends to matter more in production than in a press release.